Improved Kalman Filter Variants for UAV Tracking with Radar Motion Models

: Unmanned aerial vehicles (UAV) have made a huge inﬂuence on our everyday life with maturity of technology and more extensive applications. Tracking UAVs has become more and more signiﬁcant because of not only their beneﬁcial location-based service, but also their potential threats. UAVs are low-altitude, slow-speed, and small targets, which makes it possible to track them with mobile radars, such as vehicle radars and UAVs with radars. Kalman ﬁlter and its variant algorithms are widely used to extract useful trajectory information from data mixed with noise. Applying those ﬁlter algorithms in east-north-up (ENU) coordinates with mobile radars causes ﬁlter performance degradation. To improve this, we made a derivation on the motion-model consistency of mobile radar with constant velocity. Then, extending common ﬁlter algorithms into earth-centered earth-ﬁxed (ECEF) coordinates to ﬁlter out random errors is proposed. The theory analysis and simulation shows that the improved algorithms provide more e ﬃ ciency and compatibility in mobile radar scenes.


Introduction
Unmanned Aerial Vehicles (UAVs), also named drones, have attracted considerable attention because of their widespread use and potential threats. Aerial photography, agriculture, plant protection, express delivery, disaster relief, wildlife observation, infectious-disease monitoring, mapping, news reports, power inspection, disaster management [1], movie and television photography are common UAV applications. However, UAVs have brought many threats such as disturbing regular flight, attacking by carried dangerous goods, spying, border smuggling, and drug trafficking. It is necessary and urgent to detect and track UAVs. Computer vision [2], capturing remote control [3] and radar detection [4] are common methods to identify UAVs. The rapid development and application of internet-of-things (IoT) and fifth-generation (5G) technology prompt the economic value of UAVs across many civil application fields [5]. Many 5G-based detecting and tracking methods are invented [6]. Researchers and engineers will continue paying more attention to identify and track UAVs.
Radar automatic target recognition is a common application in radar systems. Fixed radars are commonly used to detect objects. There are also many mobile types of radars, such as vehicle-borne radars, space-borne radars, airborne radars, ship-borne radars. The carrying capacity of UAV makes it suitable to accomplish detecting mission with onboard radars. Pule Doppler and synthetic-aperture radar (SAR) are common technological regimes of radars on UAVs and radars on UAVs are used to disaster relief, movie and television photography. They have made a big difference to our daily life. However, some negative impact has also been brought in the meantime, such as disturbing the regular flights in no-fly zones, attacks by carrying dangerous objects, and spying on private property and people's privacy. They will be even more dangerous when hijacked by lawbreakers. Border smuggling, drug trafficking, and many other black flight events implemented by UAVs happen a lot and they threaten our normal life and work. Therefore, it becomes urgent and significant to probe, recognize, and track UAVs.

When UAV Meets 5G IoT
With high data rate, low latency, low power consumption, low-cost, huge system capacity, and massive device connection, fifth-generation mobile networks (5G) bring us a technical revolution and a new communication era. Its huge network capacity and connectivity for billions of devices hugely satisfy the demands of communication in the internet of things (IoT). During the gradual progress from fourth generation (4G) (Long Term Evolution-Advanced (LTE-A) [20], World Interoperability for Microwave Access (WiMAX) [21,22]) to 5G, many technologies have been invented and discussed by researchers and engineers. Authors in [23] proposed an integrated network-function virtualization with software defined radio (SDR) and software defined network (SDN) for 4G/5G networks, and discussed the extension standards and possible extensions for future mobile networks. Ultra-dense heterogeneous networks, self-organizing networks, content-distribution networks, device-to-device communication (D2D), machine-to-machine (M2M) are also key technologies in 5G. With the ever-growing demand for mobile networks, UAV application has been one of the earliest business applications in 5G. UAVs are widely used to provide a reliable and cost-effective solution to wireless communications. By connecting large numbers of base stations with UAVs, a 5G IoT network for future smart-city architecture is proposed in [24]. Their work has an obvious advantage over the existing system by simulation results. Authors in [25] designed a novel link-adaptive constellation-division multiple access (CoDMA) technique by integrating millimeter wave (mmWave) communication with UAV-aided 5G ultra-dense networks. More achievements in the integration of 5G mmWave communications with UAV-assisted wireless networks are surveyed in [26]. The tremendous data processing in 5G demands new technical improvement on cloud computing. Cloud transmission [27] in 5G will make a great difference. As a new computing model, cloud computing can provide secure and reliable data storage and convenient and quick internet service for UAVs. 5G phased arrays in base stations can be used to detect UAVs and monitor automotive traffic, which is discussed in [28]. All these different kinds of applications reveal that UAVs are playing a more and more important role in 5G IoT.

Identify by Computer Vision
By constituting a node through the deployment of a large number of cameras, and then a visual sensor network will be constituted through many nodes. The UAV's image and information storage will be captured by the whole visual sensor network. The corresponding characteristic information of UAV can be extracted from the captured computer images. After plentiful learning and training, deep learning and neural networks can accomplish recognition and classification of UAVs. But it demands a high-visibility environment. It will never work at night or in severe weather. It also costs too much by deploying huge amounts of cameras.

Detecting and Interfere UAVs by Capturing Remote Control Signals
Electromagnetic waves are the only medium of communication and remote control for UAVs. By cutting off the communication link, UAVs can be disturbed. Electromagnetic-pulse bombs and high-power microwave weapons are common tools to interfere UAVs. By means of theoretical analysis, simulation modeling, and physical verification of software radio and the electromagnetic environment, Electronics 2020, 9, 768 4 of 17 researchers did much anti-interference analysis of UAVs' radar echo. But it will not work when the UAVs switch to a silent-flight model. [15] Radar target automatic recognition technology plays an important role in radar system. There are some micromotion components on UAVs. The micro-Doppler composition could be extracted from UAVs' radar echo because the rotational motion of micromotion components on UAVs modulates the radar echo. It turns out to be periodic modulation of frequency.

Identification of UAVs Based on 5G Millimeter-Wave Cellular Infrastructure
The deployment of ubiquitous cellular networks can also be used to detect black flight UAVs. It is similar to identifying by computer vision, and it is more convenient with the aid of existing cellular base stations, although it may cause greater burdens on base stations. With the abilities of high data rate, low delay, huge system capacity, and massive device connections, 5G will make an enormous influence on our everyday life. Identification of UAVs by 5G technology will not be a problem, and researchers have done lots of research and exploration.

Kalman Filter Algorithm and Its Extended Algorithms
All filtering algorithms are based on models, which are divided into two categories, linear and nonlinear filtering algorithms. Kalman filter (KF) is one of the typical linear filtering algorithms. Extended Kalman filter (EKF), unscented Kalman filter (UKF), converted measurement Kalman filter (CMKF) and unbiased converted-measurement Kalman filter (UCMKF) are all widely used nonlinear filtering algorithms. Figure 1 depicts a typical scene of tracking UAVs There are many types of radars monitoring the airspace, such as UAV radars and vehicle radars. The alien UAV will be discovered when it enters radars' monitoring area. However, not all alien UAVs need to be attacked immediately, some of them just need to be tracked. Subject to the detecting range, the fixed radars are not capable to track UAVs all the time. Therefore, vehicle radars and UAVs with radars will be good choices to accomplish this tracking task.  3. Detecting UAVs by extracting the micro-Doppler effect [15] Radar target automatic recognition technology plays an important role in radar system. There are some micromotion components on UAVs. The micro-Doppler composition could be extracted from UAVs' radar echo because the rotational motion of micromotion components on UAVs modulates the radar echo. It turns out to be periodic modulation of frequency. 4. Identification of UAVs based on 5G millimeter-wave cellular infrastructure

A Special Application Scenario
The deployment of ubiquitous cellular networks can also be used to detect black flight UAVs. It is similar to identifying by computer vision, and it is more convenient with the aid of existing cellular base stations, although it may cause greater burdens on base stations. With the abilities of high data rate, low delay, huge system capacity, and massive device connections, 5G will make an enormous influence on our everyday life. Identification of UAVs by 5G technology will not be a problem, and researchers have done lots of research and exploration.

Kalman Filter Algorithm and Its Extended Algorithms
All filtering algorithms are based on models, which are divided into two categories, linear and nonlinear filtering algorithms. Kalman filter (KF) is one of the typical linear filtering algorithms. Extended Kalman filter (EKF), unscented Kalman filter (UKF), converted measurement Kalman filter (CMKF) and unbiased converted-measurement Kalman filter (UCMKF) are all widely used nonlinear filtering algorithms. Figure 1 depicts a typical scene of tracking UAVs There are many types of radars monitoring the airspace, such as UAV radars and vehicle radars. The alien UAV will be discovered when it enters radars' monitoring area. However, not all alien UAVs need to be attacked immediately, some of them just need to be tracked. Subject to the detecting range, the fixed radars are not capable to track UAVs all the time. Therefore, vehicle radars and UAVs with radars will be good choices to accomplish this tracking task. In order to obtain the target's locating and tracking information during tracking, we need to carry out data fusion. Measurement noise is inevitable during data acquisition and processing. It is a classical state estimation issue to extract trajectory information. Kalman Filter and its variant algorithms are popular to solve this kind of issue. Meanwhile, a variety of coordinates are employed  In order to obtain the target's locating and tracking information during tracking, we need to carry out data fusion. Measurement noise is inevitable during data acquisition and processing. It is a classical state estimation issue to extract trajectory information. Kalman Filter and its variant algorithms are popular to solve this kind of issue. Meanwhile, a variety of coordinates are employed during data processing. In this paper, we will discuss the filtering performance of EKF, UKF and UCMKF in different coordinates with mobile radars. It suits the above-mentioned scene of tracking UAVs with mobile radars. Finally, we found that filtering in east-north-up (ENU) coordinates is not adequate for scenes with mobile radars. But it works well when accomplishing filtering tasks in earth-centered earth-fixed (ECEF) coordinates, which is possible in the scene shown in Figure 1.

Analysis of Motion Model Consistency during Tracking UAVs
During tracking objects with radars, relative motion exists. The motion occurs on objects or radars, and maybe both. A motion model is introduced to describe the process. The consistency of the motion model is crucial, because it defines the validity of data processing. Coordinate transformation is an important task to keep the consistency of the motion model.

The Applied Coordinates
Coordinate system is a common auxiliary method of science. In practice, the multiformity of motion models and measuring platforms requires researchers and engineers to adopt many different kinds of coordinates in target tracking applications. Spherical coordinates, east-north-up (ENU) coordinates, earth-centered earth-fixed (ECEF) coordinates, and geodetic coordinate [29] are among widely used coordinates. The motion model consistency is emphasized much during coordinate transformation in detection and tracking. In this paper, we also discuss this issue.
Different coordinates are defined as below. Spherical coordinates: as part of a 3-D reference system, these consist of range, azimuth, and elevation. Usually, they are used to establish the measurement model [30].
ENU coordinates: these are an approximate inertial coordinate system, and use the barycenter of the platform as the origin. E is along the tangential of the earth's rotation, N is along the north, and U is perpendicular to the plane consisting of the E-axis and N-axis and points up [31]. The ENU coordinates are a local coordinate system where radars locate. Filtering is usually performed in ENU coordinates in conventional-state estimation problems [19].
ECEF coordinates: also known as the earth coordinate system, these are a spatial rectangular coordinate system based on a geographical system. The origin of ECEF coordinates coincides with the earth's center of mass, the x-axis points to the Greenwich meridian, the z-axis points along the Earth's axis, and the x-axis, y-axis and z-axis form right-handed coordinates. Usually, it is used to establish the motion model [30].
Geodetic coordinates: these are built on an ellipsoid surface based on the earth's surface. Latitude, longitude and altitude are the three coordinates. Usually, measured data are from geodetic coordinates. In our study we adopt the World Geodetic System 1984 (WGS-84) coordinates system [32].
In the traditional tracking system, the measurement model is set up in 3-D spherical coordinates and the measured data are from geodetic coordinates. Filtering is performed in ENU coordinates. It works well when radars are motionless or move with a very low velocity. This kind of radar can be applied traditional methods to track UAVs. But our work shows that when moving radars are applied to tracking UAVs, filtering random errors will not be accomplished well by traditional methods. More efficient methods to filter random errors must be developed urgently.

Problem of Interest
In this section, we present a theoretical deduction to show how filtering out random errors in ENU coordinates results in filter performance degradation when tracking UAVs with mobile radars. It disturbs the motion model consistency. In practice, the moving states of radars are complicated. The constant-velocity (CV) model, constant-turn (CT) model, and constant-acceleration (CA) model are among common motion models. The real moving state is a set of many single motion models. The CV model is taken as an example to explain motion-model consistency when filtering directly in ENU coordinates.
The state equation and measurement equation are widely used to describe systems: where X(k) denotes the state vector at time k; F denotes the state transition matrix, and usually assumed stationary; G denotes the process noise covariance matrix; Z(k) denotes the actual measurement value of X at time k; H denotes the measurement matrix; and w(k) and v(k), denote the process noise vector and measurement noise vector, respectively. Here we assume a zero mean and Gaussian white noise. Their covariance matrices are given by: In this paper, the constant-velocity (CV) model is employed to discuss the motion-model consistency in the period of coordinate transformation. Providing the radar is fastened on a gyrostabilized platform and all biases are offset, T is the sampling interval of radar, and the motion model state transition matrix will be: Then we can achieve the expression of G and H: describes the components of position and velocity of the state vector X(k). (x, y, z) T represents position components of measurement vector Z k .
Then, we analyze the consistency of the CV model during the coordinate transformation. Providing that the radar moves with a constant speed with component In the CV model, the velocity component of the target is set as (V x1 , V y1 , V z1 ) in the ECEF coordinates. After transforming to the ENU coordinates, we get the velocity (V x , V y , V z ): Electronics 2020, 9, 768 7 of 17 where, Set: V p is a constant regardless of random noise, then we get a new expression of (V x , V y , V z ): From Equation (8), we can derive that with motionless radars, there is no difference between the models in ENU coordinates and ECEF coordinates. The velocity component keeps time-invariant despite its changing value. However, when the radar moves, each element in B −1 will change, and thereby (V x , V y , V z ) varies, that means the model will no longer be a CV model while the radar moves. In this situation, the motion model's consistency is aborted. It will cause filter performance degradation during tracking.
The theoretical analysis above shows that the motion model remains unaltered when the radar is motionless. However, when the radar moves, motion models will not remain consistent after the coordinate transformation, the higher velocity, the more drastic changes will happen. The model changes will cause the filtering deviation, and then the system will never accomplish the tracking mission. Now we can conclude that performing the filtering task in ENU coordinates with CV models is not appropriate. We cannot apply ENU coordinates when tracking UAVs with mobile radars.
As Section 2.1 said, in traditional target tracking applications, the motion model is set up in ECEF coordinates and filtering is usually performed in ENU coordinates. But filtering in ENU coordinates is not practicable, which is proved by previous theoretical analysis. Therefore, extending the common filtering algorithms into ECEF coordinates is proposed to solve this issue. Figure 2 shows the flow chart of data transformation among different coordinate systems both in the traditional method and proposed method. Set: p V is a constant regardless of random noise, then we get a new expression of ( , , ) From Equation (8), we can derive that with motionless radars, there is no difference between the models in ENU coordinates and ECEF coordinates. The velocity component keeps time-invariant despite its changing value. However, when the radar moves, each element in 1 B will change, and thereby ( , , ) x y z V V V varies, that means the model will no longer be a CV model while the radar moves. In this situation, the motion model's consistency is aborted. It will cause filter performance degradation during tracking. The theoretical analysis above shows that the motion model remains unaltered when the radar is motionless. However, when the radar moves, motion models will not remain consistent after the coordinate transformation, the higher velocity, the more drastic changes will happen. The model changes will cause the filtering deviation, and then the system will never accomplish the tracking mission. Now we can conclude that performing the filtering task in ENU coordinates with CV models is not appropriate. We cannot apply ENU coordinates when tracking UAVs with mobile radars.
As Section 2.1 said, in traditional target tracking applications, the motion model is set up in ECEF coordinates and filtering is usually performed in ENU coordinates. But filtering in ENU coordinates is not practicable, which is proved by previous theoretical analysis. Therefore, extending the common filtering algorithms into ECEF coordinates is proposed to solve this issue. Figure 2 shows the flow chart of data transformation among different coordinate systems both in the traditional method and proposed method.

Extended Algorithms Based on ECEF Coordinates
As mentioned before, filter deviation and even divergence will result from the model inconsistency. This study proposed to carry out the filtering in the ECEF coordinates to ensure the motion model's consistency, and extend common filtering algorithms from the ENU coordinates to the ECEF coordinates.
As mentioned, the measurement data ( r, θ, η) are achieved from the spherical coordinates. (r, θ, η) are the practical measurement data, (r σ , θ σ , η σ ) are the random measurement errors, and (σ r , σ θ , σ η ) are the standard deviation of measurement, which features a zero-mean Gaussian distribution. Registration is applied to eliminate the system errors. (L(k), λ(k), H(k)) and (x s (k), y s (k), z s (k)) are position components of the radar in the geodetic coordinates and ECEF coordinates, respectively.

EKF Algorithm Based on ECEF Coordinates
When filtering is performed in the ENU coordinates, the state equations are set in the space rectangular coordinates (namely the ENU coordinates). The radar is located in the ENU coordinates, but the measurement equations are set in the spherical coordinates. The state equation is a linear equation and it is the same as the state equation in (1). The measurement equation is a nonlinear equation. It is expressed as follows: The form of h[X(k)] is (A1) showed in Appendix A. It is different from linear systems, the state transition matrix and measurement matrix in nonlinear systems cannot be expressed directly in the form of matrix, but in the form of f [X(k)] and h[X(k)]. In order to linearize nonlinear functions, EKF use Taylor expansion in the neighborhood of mean value of the state variable. The measurement matrix H can be achieved by calculating the Jacobian matrix of h[X(k)], as shown in (A2) in Appendix A, where X(k k − 1) means the state vector at time k, which is predicted from time k − 1. r = x 2 + y 2 + z 2 ,r 1 = x 2 + y 2 .
After calculating H, the format is similar to the measurement matrix in the linear system, and the filtering task can be accomplished by Kalman filter.
When applying and extending the application of the EKF from the ENU coordinate to the ECEF coordinate, the calculation of the Jacobian matrix is too complicated to be employed in an engineering application. As the state equation is set in the spherical coordinates, the measurement data have to be converted from ECEF to spherical coordinates. This calculation is too large to consider in this paper. We just show the performance degradation when applying EKF directly in ENU coordinates.

UKF Algorithm Based on ECEF Coordinates
EKF linearizes the nonlinear equations by the first-order Taylor expansion, and then adopts the Kalman filter to accomplish the filtering task. However, the divergence of the approximation and huge calculating burden of the Jacobian matrix will make EKF insignificant in some applications. Therefore, an unscented conversion based on the unscented Kalman filter algorithm was proposed in [31].
UKF calculates the assigned σ sample point through an unscented conversion method to linearize the nonlinear equation. The measurement sampling and weight of the σ sampling point determine the final measurement prediction. The heavily computation of the Jacobian matrix is avoided by this procedure. In Reference [31], the UKF algorithm based on the ENU coordinates was discussed.
After being calculated from the ECEF coordinate, the state equation will be set up in the geocentric coordinates. To obtain the measurement vector's σ sampling, we calculate the sampling point in the ECEF coordinates first and then convert it into the ENU coordinates system. At last, the sampling of a one-step prediction will be determined. The following three steps describe the conversion process:

1.
Calculating the σ sampling of the measurement vector: here, n means the dimension of state X k , n equals 6, and λ determines the dispersive degree of the sampling point. It is usually set as a small positive value (for example, 0.01). k is usually set as 0.
(n + λ)P k|k−1 i denotes line i of the square-rooting matrix of matrix (n + λ)P k|k−1 .
This step can be divided in four steps, which are shown in Equations (A3)-(A7) in Appendix A.

2.
Converting the sample point into the ENU coordinate: , ξ i (k)(5) denote the first, third, and fifth components of ξ i (k).

3.
Calculating the measurement prediction: After conversion, the filtering task will be accomplished by the common UKF.

UCMKF Algorithm Based on ECEF Coordinates
First of all, we need to convert from the spherical coordinates to the ENU coordinates: x = (r + r σ ) cos(η + η σ ) cos(θ + θ σ ) y = (r + r σ ) cos(η + η σ ) sin(θ + θ σ ) z = (r + r σ ) sin(η + η σ ) (14) In the course of above data conversion, the coupling of each component caused by random measurement error will weaken system performance. It is urgent to remove the coupling bias. The authors of [33] developed an unbiased converted-measurement Kalman filter (UCMKF) algorithm in 3-D space to amend the conversion measurements. The corrected measurements (x u , y u , z u ) and covariance matrix R u are given: x u = r cos η cos θ(1 + λ θ λ η − λ −1 The elements of matrix R u are given by (A8) in Appendix A.
(x u , y u , z u ) is the target's corrected position in the ENU coordinates. We transform measurement data into the ECEF coordinates to guarantee the motion model consistency. The target's position (x t , y t , z t ) in the ECEF coordinates will be: The measurement covariance matrix R p in the ECEF coordinates is achieved according to properties of covariance: Converting the data measurement and the covariance matrix into the ECEF coordinate ensures the motion model's consistency. Therefore, the standard Kalman filter can be applied to filter out the measured data after the conversion; Reference [10] provides the related formula.

Simulation Experiment and Data Analysis
MATLAB simulation software is a commercial mathematics software with excellent performance in developing algorithms, analyzing data, modeling, and simulating. It was chosen to testify the validity of our proposed algorithms in this study. As mentioned, the calculation of the Jacobian matrix is too complicated to apply the EKF algorithm in the ECEF coordinate, and we only simulate the EKF algorithm in the ENU coordinates. In addition, we also simulate UKF and UCMKF in the ENU coordinate and ECEF coordinate to compare their filtering performance.

Experiment Setup
In reality, we track UAVs with radars installed on vehicles or UAVs. For ease of expression, in our simulation, moving with a CV model, the radar tracks a flying UAV in the sky. Its initial geographical coordinates are [N39.9 • , E116.4 • , 100 m], and geographical coordinates at time k are given by: We choose velocity coefficient v to simulate the real velocity of the moving platform. Here v equals 0, 1, 10, and 20, which means the platform is static, moving with low velocity and two higher velocities. We set the radar's sampling interval as 1s, and the standard deviations of the random measurement noises as [100 m, 0.08 • , 0.08 • ]. Registration is already implemented to remove the offset and attitude biases. The initial target's coordinates are [−2921 km, 4726 km, 3120 km] in ECEF coordinates.
The Monte Carlo method is one numerical computation method based on probability and statistics theory, and it can obtain numerical results through large numbers of random repeated sampling.
In this study, the simulation step is set as 50, and the Monte Carlo run number as 100. The root-mean-square error (RMSE) at time k is: Root-mean-square error of position: (22) x(k), y(k), z(k) and x(k), y(k), z(k) denote the real position components and the position components after filtering along x, y, z-axis at time k. N is the Monte Carlo simulation number. We evaluate the filter capacity by the root-mean-square error (RMSE), the smaller it is, the better filter capacity.

Simulation Results and Discussion
The results of simulation are presented in Figures 3-6. The black lines with rhombuses depict the raw data's location RMSEs. The green lines with dots show the filtered data's location RMSEs in the ECEF coordinates. The blue lines with triangles describe the filtered data's location RMSEs in the ENU coordinates. The detailed quantized data are displayed in Table 1.
x k x k y k y k z k z k (22) ( ), ( ), ( ) x k y k z k and ( ), ( ), ( ) x k y k z k denote the real position components and the position components after filtering along ,, xyz-axis at time k . N is the Monte Carlo simulation number. We evaluate the filter capacity by the root-mean-square error (RMSE), the smaller it is, the better filter capacity.

Simulation Results and Discussion
The results of simulation are presented in Figures 3-6. The black lines with rhombuses depict the raw data's location RMSEs. The green lines with dots show the filtered data's location RMSEs in the ECEF coordinates. The blue lines with triangles describe the filtered data's location RMSEs in the ENU coordinates. The detailed quantized data are displayed in Table 1.           Figure 3, i.e., the radar is static. It is with the situation to tracking UAVs with fixed radars. The filterings by EKF, UKF, and UCMKF all work well enough in the ENU coordinates and these are almost the same as those in ECEF coordinates, as shown in Table 1. Figure 4 depicts the simulation results when the motion coefficient is set as 1, which means the radar is moving slowly. Here, the filtering results by EKF, UKF, and UCMKF show no obvious deviations in the ENU coordinate, which is shown in Figure 4b,c. The filtering results in ECEF coordinates show no significant differences from those in ENU coordinates. However, the data analysis in Table 1 shows that the filtering accuracy degrades slightly, when filtering is performed in the ECEF coordinates. This is because some errors arise during the coordinate system conversion. In these cases, the traditional method works well and there is no need to worry about the filtering performance degradation. Figures 5 and 6 show the filtering results in the case of moving radars with high speed. It corresponds to the special situation mentioned in Section 2. As described in Figure 5, when the   6 show the filtering performance with the different velocity coefficient of radars. The velocity coefficient is 0 in Figure 3, i.e., the radar is static. It is with the situation to tracking UAVs with fixed radars. The filterings by EKF, UKF, and UCMKF all work well enough in the ENU coordinates and these are almost the same as those in ECEF coordinates, as shown in Table 1. Figure 4 depicts the simulation results when the motion coefficient is set as 1, which means the radar is moving slowly. Here, the filtering results by EKF, UKF, and UCMKF show no obvious deviations in the ENU coordinate, which is shown in Figure 4b,c. The filtering results in ECEF coordinates show no significant differences from those in ENU coordinates. However, the data analysis in Table 1 shows that the filtering accuracy degrades slightly, when filtering is performed in the ECEF coordinates. This is because some errors arise during the coordinate system conversion. In these cases, the traditional method works well and there is no need to worry about the filtering performance degradation. Figures 5 and 6 show the filtering results in the case of moving radars with high speed. It corresponds to the special situation mentioned in Section 2. As described in Figure 5, when the motion coefficient is 10, the filtering performance in the ENU coordinates degrades. However, Figure 5b,c show that the UKF and UCMKF algorithms in the ECEF coordinates work as well as usual. When the motion coefficient goes up to 20-in other words, the radar moves more quickly-the filtering divergence occurs via the EKF, UKF, and UCMKF algorithms in the ENU coordinates. The data are no longer efficient because their accuracy is lower than the raw data, as shown in Figure 6. This performance degradation will never be accepted. However, filtering in the ECEF coordinates still works well. The proposed method is efficient when tracking UAVs with moving radars.
We can draw a conclusion that in a low-speed CV model (targets and radars move at a constant velocity) the filtering results in both the ENU and ECEF coordinates are efficient. These two methods are both capable for tracking UAVs in this case. When the radar moves quickly enough the filtering degradation and result divergence will occur in the ENU coordinates. Tracking UAVs will fail by this means. We need to filter data in the ECEF coordinates, because that provides more precise filter results no matter how fast the radar moves. The simulation results fit well with the previous theoretical analysis. The proposed extension algorithms have been proved effective by simulation.

Computational Complexity and Implementation Cost of Proposed Extended Algorithms
In this section, we will discuss the quality of the proposed extended algorithm. The validity has been approved by simulation results. The advantages of proposed extended algorithms are necessary to investigate.

Computational Complexity
The time complexity of EKF is O(n 2 ) when process noise and measurement noise feature Gaussian distribution. As mentioned before, the calculation of the Jacobian matrix is too complicated to be employed in an engineering application. We will not consider that in this paper. However, we calculate the time complexity after checking our code. The extended UKF algorithm and UCMKF algorithms are O(n 3 ) and O(n). The time complexity of UKF is high, but the amount of code is little, so the run time is not long.
In our paper, we chose MATLAB software to simulate the whole calculation of proposed extended algorithms. The simulation ran on a personal laptop. The detailed hardware configurations are Intel Core i3-2310M, 2.1 GHz, 4 GB RAM, and a 120 GB SSD. The total running times on this computer for three algorithms are 2 s, 5.2 s, and 2.2 s. The simulation is performed on a laptop, whose performance is not strong enough. When performing this simulation on a desk computer with a more powerful configuration, the run time will be extremely reduced. When applying our extended algorithms on UAVs with radars and vehicle radars, it will cause little burden for radars. It is a lightweight task.

Implementation Cost
In order to track UAVs with mobile radars, we proposed extended filtering algorithms with short running time. Vehicle radars are common after a long development history. The existing vehicle radars can be easily equipped with our algorithms to improve UAV-tracking performance while moving. The implementation cost is not high at all. As for UAVs with radars, Pule Doppler and synthetic-aperture radar (SAR) are common technological regimes of radars on UAVs. Nowadays, the UAVs become cheap enough to perform scaled applications. Radars are also mature products and cheap enough to afford. The computational and communication ability of UAVs are high enough to perform our proposed extended algorithms. The implementation cost is also low enough.

Scalability Analysis
The scalability of our proposed method is good. However, in our proposed architecture, we only considered a single radar to track UAVs. It weakens the continuous tracking ability and system reliability. To improve the system performance, we will develop a multi-UAV cooperative application to form an ad hoc network. The vehicle radars can also connect to the UAV networks to make a supplement. The system communication ability can be guaranteed with the help of LTE, 5G, Wi-Fi and other advanced communication technology. The proposed method has potential in large-scale scenarios.

Conclusions
In this article, a special situation of tracking UAVs with mobile radars is discussed. We found that adopting EKF, UKF, and UCMKF directly in the ENU coordinates will cause filter-performance degradation or even divergence with mobile radars by theoretical motivation. That will degrade tracking performance of tracking UAVs with mobile radars, especially when radars move quickly. A new method is proposed to perform the filtering operation in the ECEF coordinates. The theory analysis and simulation results show that the proposed method can unify the motion model and guarantee the accuracy of the filtering process. With low cost and good scalability, it can constitute a UAV network in the future to improve the system performance. It is helpful to avert filtering performance degradation, and even divergence, when tracking UAVs with mobile radars. In the future, more motion models such as the constant-turn model, constant-acceleration model, and variable-acceleration models will be focused on to make it more reliable and practical when tracking UAVs. It is also significant in engineering applications. Acknowledgments: Partial results of this work have been published in IIKI 2018. The authors would like to thank the helpful comments of the anonymous reviewers.

Conflicts of Interest:
The authors declare no conflict of interest.

Appendix A
In this part, we present some well-known equations to improve the readability of our paper.
The calculation of the σ sampling of the measurement vector are as follows.